Research map / Adversarial robustness and security / Certified and empirical robustness
Holistic Adversarially Robust Pruning
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Jointly learn layer-wise compression quotas and connection importance scores (ablation) | HARP · CIFAR-10 and SVHN, 99% and 99.9% sparsity | natural accuracy and adversarial robustness | Helped |
| Add learnable layer-wise compression quotas (ablation) | HARP · CIFAR-10, ResNet-18 and VGG | natural accuracy and PGD-10 robustness | Helped |
| Add learnable connection importance scores (ablation) | HARP · CIFAR-10, ResNet-18 and VGG | natural accuracy and PGD-10 robustness | Helped |
| Initialize layer compression rates non-uniformly with ERK or LAMP rather than uniformly (ablation) | HARP · CIFAR-10, 99% and 99.9% sparsity | natural accuracy and PGD-10 accuracy: 80.25; 50.36 -> 79.42; 50.43; 80.48; 50.33 | No clear effect |
| Use a dynamic regularization step size of gamma=0.01 rather than smaller or larger step sizes (ablation) | HARP · VGG, 99.9% sparsity | convergence | Helped |
| Reduce the dynamic regularization step size to gamma=0.001 (ablation) | HARP · VGG, 99.9% sparsity | target compression reached | Hurt |
| Increase the dynamic regularization step size to gamma=1.0 (ablation) | HARP · VGG, 99.9% sparsity | performance improvement | Hurt |
| Use non-uniform LAMP or ERK layer rates with Hydra instead of uniform compression | Hydra · CIFAR-10, 99% and 99.9% sparsity | natural accuracy and adversarial robustness | Helped |
| Use LAMP or ERK layer rates with Hydra rather than its original strategy | Hydra · ImageNet, 99% sparsity | model performance | Hurt |
| Prune using HARP rather than Hydra | robustly pretrained models · CIFAR-10, SVHN, and ImageNet, especially at 99.9% sparsity | natural accuracy and adversarial robustness | Helped |
| Prune using HARP rather than R-ADMM | robustly pretrained models · CIFAR-10, SVHN, and ImageNet | natural accuracy and adversarial robustness: 20 percentage points for natural accuracy; 10 percentage points for robustness | Helped |
| Use HARP rather than MAD | robustly pretrained models · CIFAR-10 and SVHN at 99% sparsity | PGD-10 robustness | Helped |
Closest papers by meaning
- “Learning-Compression” Algorithms for Neural Net Pruning (CVPR 2018 · similarity 0.90)
- Data-Independent Neural Pruning via Coresets (ICLR 2020 · similarity 0.89)
- One Less Reason for Filter Pruning: Gaining Free Adversarial Robustness with Structured Grouped Kernel Pruning (NeurIPS 2023 · similarity 0.89)
- Adaptive Sharpness-Aware Pruning for Robust Sparse Networks (ICLR 2024 · similarity 0.88)
- Model Compression with Adversarial Robustness: A Unified Optimization Framework (NeurIPS 2019 · similarity 0.88)
- Dynamical Low-Rank Compression of Neural Networks with Robustness under Adversarial Attacks (NeurIPS 2025 · similarity 0.87)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.